@ag-bash/mcp-server

@ag-bash/mcp-server

A standalone Model Context Protocol (MCP) server that provides a persistent, sandboxed Bash environment to AI agents.

This server leverages the @ag-bash/bash engine to allow agents to execute shell commands, process data with jq, and run Python/JS scripts within a secure virtual filesystem.

🚀 Installation

Install the MCP server globally using npm:

npm install -g @ag-bash/mcp-server

⚙️ Configuration

To use Ag-Bash with your favorite AI client, add it to your mcpConfig.json or equivalent configuration file.

Claude Desktop

On macOS, edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "ag-bash": {
      "command": "ag-bash-mcp",
      "args": []
    }
  }
}

Cursor / VS Code

Add a new MCP server in your IDE settings with the following details:

• Name: ag-bash
• Type: command
• Command: ag-bash-mcp

🛠️ Available Tools

The server exports the following tools to the agent:

1. run_bash

Executes a bash script in the sandboxed environment.

• Parameters:
  • script (string, required): The bash code to execute.
• Features:
  • Automatically persists environment variables, functions, and CWD across calls.
  • Full access to jq, grep, sed, and other unix utilities.

2. get_state

Retrieves the current state of the shell session.

• Returns:
  • Current working directory (cwd).
  • Active environment variables.
  • Defined functions.

🛡️ Security Features

• Read-Only Root: By default, the server mounts your current project root as a Read-Only overlay. Any changes made by the agent stay in the virtual memory and never touch your real files.
• In-Process Sandbox: No external VM required; execution is isolated using Ag-Bash's internal security logic.
• Resource Limits: Protects against infinite loops and excessive memory consumption.

📜 License

Apache-2.0